Skip to content

Latest commit

 

History

History
341 lines (288 loc) · 15.5 KB

trtllm_guide.md

File metadata and controls

341 lines (288 loc) · 15.5 KB
Raw

Deploying Hugging Face Llama2-7b Model in Triton

TensorRT-LLM is Nvidia's recommended solution of running Large Language Models(LLMs) on Nvidia GPUs. Read more about TensoRT-LLM here and Triton's TensorRT-LLM Backend here.

NOTE: If some parts of this tutorial doesn't work, it is possible that there are some version mismatches between the tutorials and tensorrtllm_backend repository. Refer to llama.md for more detailed modifications if necessary. And if you are familiar with python, you can also try using High-level API for LLM workflow.

Acquiring Llama2-7B model

For this tutorial, we are using the Llama2-7B HuggingFace model with pre-trained weights. Clone the repo of the model with weights and tokens here. You will need to get permissions for the Llama2 repository as well as get access to the huggingface cli. To get access to the huggingface cli, go here: huggingface.co/settings/tokens.

Deploying with Triton CLI

Triton CLI is an open source command line interface that enables users to create, deploy, and profile models served by the Triton Inference Server.

Launch Triton TensorRT-LLM container

Launch Triton docker container with TensorRT-LLM backend. Note that we're mounting the acquired Llama2-7b model to /root/.cache/huggingface in the docker container so that Triton CLI could use it and skip the download step.

Make an engines folder outside docker to reuse engines for future runs. Please, make sure to replace <xx.yy> with the version of Triton that you want to use.

docker run --rm -it --net host --shm-size=2g \
    --ulimit memlock=-1 --ulimit stack=67108864 --gpus all \
    -v </path/to/Llama2/repo>:/root/.cache/huggingface \
    -v </path/to/engines>:/engines \
    nvcr.io/nvidia/tritonserver:<xx.yy>-trtllm-python-py3

Install Triton CLI

Install the latest release of Triton CLI:

GIT_REF=<LATEST_RELEASE>
pip install git+https://github.com/triton-inference-server/triton_cli.git@${GIT_REF}

Prepare Triton model repository

Triton CLI has a single command triton import that automatically converts HF checkpoint into TensorRT-LLM checkpoint format, builds TensorRT-LLM engines, and prepares a Triton model repository:

ENGINE_DEST_PATH=/engines triton import -m llama-2-7b --backend tensorrtllm

Please, note that specifying ENGINE_DEST_PATH is optional, but recommended if you want to re-use compiled engines in the future.

After successful run of triton import, you should see the structure of a model repository printed in the console:

...
triton - INFO - Current repo at /root/models:
models/
├── llama-2-7b/
│   ├── 1/
│   │   ├── lib/
│   │   │   ├── decode.py
│   │   │   └── triton_decoder.py
│   │   └── model.py
│   └── config.pbtxt
├── postprocessing/
│   ├── 1/
│   │   └── model.py
│   └── config.pbtxt
├── preprocessing/
│   ├── 1/
│   │   └── model.py
│   └── config.pbtxt
└── tensorrt_llm/
    ├── 1/
    └── config.pbtxt

Start Triton Inference Server

Start server pointing at the default model repository:

triton start

Send an inference request

Use the generate endpoint. to send an inference request to the deployed model.

curl -X POST localhost:8000/v2/models/llama-2-7b/generate -d '{"text_input": "What is ML?", "max_tokens": 50, "bad_words": "", "stop_words": "", "pad_id": 2, "end_id": 2}'

You should expect the following response:

{"context_logits":0.0,...,"text_output":"What is ML?\nML is a branch of AI that allows computers to learn from data, identify patterns, and make predictions. It is a powerful tool that can be used in a variety of industries, including healthcare, finance, and transportation."}

Deploying with Triton Inference Server

If you would like to hava a better control over the deployment process, next steps will guide you over the process of TensorRT-LLM engine building process and Triton model repository set up.

Prerequisite: TensorRT-LLM backend

This tutorial requires TensorRT-LLM Backend repository. Please note, that for best user experience we recommend using the latest release tag of tensorrtllm_backend and the latest Triton Server container.

To clone TensorRT-LLM Backend repository, make sure to run the following set of commands.

git clone https://github.com/triton-inference-server/tensorrtllm_backend.git  --branch <release branch>
# Update the submodules
cd tensorrtllm_backend
# Install git-lfs if needed
apt-get update && apt-get install git-lfs -y --no-install-recommends
git lfs install
git submodule update --init --recursive

Launch Triton TensorRT-LLM container

Launch Triton docker container with TensorRT-LLM backend. Note that we're mounting tensorrtllm_backend to /tensorrtllm_backend and the Llama2 model to /Llama-2-7b-hf in the docker container for simplicity. Make an engines folder outside docker to reuse engines for future runs. Please, make sure to replace <xx.yy> with the version of Triton that you want to use.

docker run --rm -it --net host --shm-size=2g \
    --ulimit memlock=-1 --ulimit stack=67108864 --gpus all \
    -v </path/to/tensorrtllm_backend>:/tensorrtllm_backend \
    -v </path/to/Llama2/repo>:/Llama-2-7b-hf \
    -v </path/to/engines>:/engines \
    nvcr.io/nvidia/tritonserver:<xx.yy>-trtllm-python-py3

Alternatively, you can follow instructions here to build Triton Server with Tensorrt-LLM Backend if you want to build a specialized container.

Don't forget to allow gpu usage when you launch the container.

Optional: For simplicity, we've condensed all following steps into a deploy_trtllm_llama.sh. Make sure to clone tutorials repo to your machine and start the docker container with the tutorial repo mounted to /tutorials by adding -v /path/to/tutorials/:/tutorials to docker run command, listed above. Then, when container has started, simply run the script via

/tutorials/Popular_Models_Guide/Llama2/deploy_trtllm_llama.sh <WORLD_SIZE>

For how to run an inference request, refer to the Client section of this tutorial.

Create Engines for each model [skip this step if you already have an engine]

TensorRT-LLM requires each model to be compiled for the configuration you need before running. To do so, before you run your model for the first time on Triton Server you will need to create a TensorRT-LLM engine.

Starting with 24.04 release, Triton Server TensrRT-LLM container comes with pre-installed TensorRT-LLM package, which allows users to build engines inside the Triton container. Simply follow the next steps:

HF_LLAMA_MODEL=/Llama-2-7b-hf
UNIFIED_CKPT_PATH=/tmp/ckpt/llama/7b/
ENGINE_DIR=/engines
CONVERT_CHKPT_SCRIPT=/tensorrtllm_backend/tensorrt_llm/examples/llama/convert_checkpoint.py
python3 ${CONVERT_CHKPT_SCRIPT} --model_dir ${HF_LLAMA_MODEL} --output_dir ${UNIFIED_CKPT_PATH} --dtype float16
trtllm-build --checkpoint_dir ${UNIFIED_CKPT_PATH} \
            --remove_input_padding enable \
            --gpt_attention_plugin float16 \
            --context_fmha enable \
            --gemm_plugin float16 \
            --output_dir ${ENGINE_DIR} \
            --paged_kv_cache enable \
            --max_batch_size 4

Optional: You can check test the output of the model with run.py located in the same llama examples folder.

 python3 /tensorrtllm_backend/tensorrt_llm/examples/run.py --engine_dir=/engines/1-gpu/ --max_output_len 50 --tokenizer_dir /Llama-2-7b-hf --input_text "What is ML?"

You should expect the following response:

[TensorRT-LLM] TensorRT-LLM version: 0.9.0
...
[TensorRT-LLM][INFO] Max KV cache pages per sequence: 1
Input [Text 0]: "<s> What is ML?"
Output [Text 0 Beam 0]: "
ML is a branch of AI that allows computers to learn from data, identify patterns, and make predictions. It is a powerful tool that can be used in a variety of industries, including healthcare, finance, and transportation."

Serving with Triton

The last step is to create a Triton readable model. You can find a template of a model that uses inflight batching in tensorrtllm_backend/all_models/inflight_batcher_llm. To run our Llama2-7B model, you will need to:

  1. Copy over the inflight batcher models repository
cp -R /tensorrtllm_backend/all_models/inflight_batcher_llm /opt/tritonserver/.
  1. Modify config.pbtxt for the preprocessing, postprocessing and processing steps. The following script do a minimized configuration to run tritonserver, but if you want optimal performance or custom parameters, read details in documentation and perf_best_practices:
# preprocessing
TOKENIZER_DIR=/Llama-2-7b-hf/
TOKENIZER_TYPE=auto
DECOUPLED_MODE=false
MODEL_FOLDER=/opt/tritonserver/inflight_batcher_llm
MAX_BATCH_SIZE=4
INSTANCE_COUNT=1
MAX_QUEUE_DELAY_MS=10000
FILL_TEMPLATE_SCRIPT=/tensorrtllm_backend/tools/fill_template.py
python3 ${FILL_TEMPLATE_SCRIPT} -i ${MODEL_FOLDER}/preprocessing/config.pbtxt tokenizer_dir:${TOKENIZER_DIR},tokenizer_type:${TOKENIZER_TYPE},triton_max_batch_size:${MAX_BATCH_SIZE},preprocessing_instance_count:${INSTANCE_COUNT}
python3 ${FILL_TEMPLATE_SCRIPT} -i ${MODEL_FOLDER}/postprocessing/config.pbtxt tokenizer_dir:${TOKENIZER_DIR},tokenizer_type:${TOKENIZER_TYPE},triton_max_batch_size:${MAX_BATCH_SIZE},postprocessing_instance_count:${INSTANCE_COUNT}
python3 ${FILL_TEMPLATE_SCRIPT} -i ${MODEL_FOLDER}/tensorrt_llm_bls/config.pbtxt triton_max_batch_size:${MAX_BATCH_SIZE},decoupled_mode:${DECOUPLED_MODE},bls_instance_count:${INSTANCE_COUNT}
python3 ${FILL_TEMPLATE_SCRIPT} -i ${MODEL_FOLDER}/ensemble/config.pbtxt triton_max_batch_size:${MAX_BATCH_SIZE}
python3 ${FILL_TEMPLATE_SCRIPT} -i ${MODEL_FOLDER}/tensorrt_llm/config.pbtxt triton_max_batch_size:${MAX_BATCH_SIZE},decoupled_mode:${DECOUPLED_MODE},engine_dir:${ENGINE_DIR},max_queue_delay_microseconds:${MAX_QUEUE_DELAY_MS},batching_strategy:inflight_fused_batching
  1. Launch Tritonserver

Use the launch_triton_server.py script. This launches multiple instances of tritonserver with MPI.

python3 /tensorrtllm_backend/scripts/launch_triton_server.py --world_size=<world size of the engine> --model_repo=/opt/tritonserver/inflight_batcher_llm

You should expect the following response:

...
I0503 22:01:25.210518 1175 grpc_server.cc:2463] Started GRPCInferenceService at 0.0.0.0:8001
I0503 22:01:25.211612 1175 http_server.cc:4692] Started HTTPService at 0.0.0.0:8000
I0503 22:01:25.254914 1175 http_server.cc:362] Started Metrics Service at 0.0.0.0:8002

To stop Triton Server inside the container, run:

pkill tritonserver

Send an inference request

You can test the results of the run with:

  1. The inflight_batcher_llm_client.py script.
# Using the SDK container as an example
docker run --rm -it --net host --shm-size=2g \
    --ulimit memlock=-1 --ulimit stack=67108864 --gpus all \
    -v /path/to/tensorrtllm_backend/inflight_batcher_llm/client:/tensorrtllm_client \
    -v /path/to/Llama2/repo:/Llama-2-7b-hf \
    nvcr.io/nvidia/tritonserver:<xx.yy>-py3-sdk
# Install extra dependencies for the script
pip3 install transformers sentencepiece
python3 /tensorrtllm_client/inflight_batcher_llm_client.py --request-output-len 50 --tokenizer-dir /Llama-2-7b-hf/ --text "What is ML?"

You should expect the following response:

...
Input: What is ML?
Output beam 0:
ML is a branch of AI that allows computers to learn from data, identify patterns, and make predictions. It is a powerful tool that can be used in a variety of industries, including healthcare, finance, and transportation.
...
  1. The generate endpoint.
curl -X POST localhost:8000/v2/models/ensemble/generate -d '{"text_input": "What is ML?", "max_tokens": 50, "bad_words": "", "stop_words": "", "pad_id": 2, "end_id": 2}'

You should expect the following response:

{"context_logits":0.0,...,"text_output":"What is ML?\nML is a branch of AI that allows computers to learn from data, identify patterns, and make predictions. It is a powerful tool that can be used in a variety of industries, including healthcare, finance, and transportation."}

References

For more examples feel free to refer to End to end workflow to run llama.